Detection of Bughole on Concrete Surface with Convolutional Neural Network

Abstract

Bugholes are surface imperfections found on the surface of concrete structures. The presence of bugholes not only affects the appearance of the concrete structure, but may even affect the durability of the structure. Traditional measurement methods are carried out by in-situ manual inspection, and the detection process is time-consuming and difficult. Although various image processing technologies (IPT) have been implemented to detect defects in the appearance quality of concrete to partially replace manual on-site inspections, the wide variety of realities may limit the widespread adoption of IPTs. In order to overcome these limitations, this paper proposes a detector based on Convolutional Neural Network (CNN) to recognizing bugholes on concrete surfaces. The proposed CNN was trained on 4,000 images and tested on 800 images which were not used for training and validation; the recognition accuracy reached 94.37%. The image test results and comparative study with traditional methods showed that the proposed method exhibits excellent performance and indeed can detect the bugholes on the concrete surfaces under actual conditions.